# Compact evidence plan, specified before native generation

Prepare a reusable evidence bundle only after queue COMPLETE, the frozen audit
COMPLETE_AUDITED, and descriptive reporting. This is not a publication decision.

Include all extracted native candidates and correctness vectors, all extracted
training candidates/rewards with identity, length and EOS status, frozen code and
contract/split, complete source-retention targets, all summaries, work logs,
interruption record, and the integrated report. Retain raw-original hashes.
A CPU checker can regrade every extracted candidate and reconstruct native
contrasts/gate plus training exposure and optimizer-update counts. This is not
independent re-extraction from every full response or optimizer replay.

Fixed full-text sample selection: first physical record in every training chunk
(one draw each;16 total) and first physical record in every native chunk (10
records total), with tokens and decoded text. No selection by correctness,
length, topic, arm performance or mechanism interpretation. This small sample
supports independent re-extraction only for those records; all other full
trajectories stay in the research workspace with hashes. Explain the limitation.

The supplement's compressed AND expanded sizes must be strictly below2,000,000
bytes. If necessary use transparent table column schemas and remove duplicated
summaries/metadata, never omit unfavorable outcomes or shrink the fixed raw
sample after inspecting it. A revised paper can link the prior paper12v1
supplement rather than duplicating its evidence.
